2026

Solaar

A multi-agent development agency built entirely with Claude Code. One orchestrator, 24 specialized agents, and an enforcement layer that makes its own rules unbreakable — the system knows your stack history, your production state, and which structural patterns in generated code are likely to fail before they reach users.

Visit Solaar
Stack
  • CLAUDE CODE
  • AGENTS
  • MCP
  • GITHUB
  • CLAUDE DESIGN

What it does

  • Accepts a project brief and runs a sequence drawn from 24 specialized agents — strategist, creative director, UX designer, frontend, mobile frontend, backend, security, QA, CI/CD, monitoring, PM, architect review, and production sentinel among them
  • Strategist selects one of eight pipeline archetypes at Step 0 — landing page, SaaS web, API service, mobile Expo, mobile native Android, content site, AI app, or web-plus-app — and skips steps that don't apply
  • Creative Director generates three divergent creative directions before any design work starts; the Architect runs a feasibility scan on each, and a blocking checkpoint locks the direction before UX touches it
  • UX agent operationalizes the chosen direction into a full design spec: color palette, typography scale, wireframes per route, responsive breakpoints, and an accessibility checklist
  • Frontend and backend agents run in parallel once the spec is approved — they write production-ready code, not scaffolds
  • Architect Review agent (Step 4.5) checks structural correctness before security: deployment target conformance, dual code path detection, API contract alignment, env var completeness, and Danger Map violations — findings classified as BLOCK, WARN, or INFO
  • Security agent audits output against OWASP Mobile/Web Top 10 before handoff; PM agent writes the release plan and task checklist
  • Production Sentinel fetches Vercel deployment status, Sentry error counts, and Lighthouse scores post-deploy; session open reads production health before any feature work starts
  • Client Knowledge Base tracks stack history, approved libraries, and accumulated lessons per client across all projects — Strategist reads it at Step 0 before any recommendation
  • 19 behavioral rules, each written from a real production incident, are enforced mechanically by Claude Code hooks — not by asking the model to remember them

Why it matters

  • Compresses time-to-first-deployable-repo from days to under 4 hours — strategy, design, frontend, backend, security, and PM done in one pipeline run
  • Blocking checkpoints prevent the most common AI coding failure mode: generating code before the design is validated
  • Prose in an instruction file does not change agent behavior — a gate does. Every rule that kept getting violated got a hook, and a regression scenario that proves the hook still works
  • The system opens every session with production awareness — errors, performance regressions, and failed deploys surface before any new feature work begins
  • Ships to Vercel automatically via GitHub Actions — no manual deploy steps after the pipeline runs

Architecture

  • Claude Code as the runtime — each agent is a markdown instruction file read and executed in sequence
  • Eight pipeline archetypes map to specific step sets — a landing page skips the architect review and backend entirely; a full SaaS runs everything
  • Blocking checkpoints between agents require explicit user approval — strategy before creative direction, direction before design, design before code
  • Claude Code hooks enforce the rules the model kept breaking: a routing gate blocks source edits when the owning specialist agent was never invoked, a PR-state gate refuses pushes to merged branches, a secret-read gate blocks .env reads before they happen, and a background-claim gate blocks any response claiming work is running when no task was launched
  • Claude Design runs between the UX and Frontend agents to generate visual prototypes from the text spec before implementation
  • Project Danger Map captures the 3 landmines, 2 unconventional conventions, and 1 non-negotiable constraint from each session; every agent reads recent entries before starting
  • GitHub Actions CI/CD and Vercel Git Integration handle build, test, and deploy on every push to main

Philosophy

  • Describe the project once — the agents handle the rest, in the right order
  • Blocking checkpoints keep the user in control without being in the loop on every decision
  • Every rule came from an incident that already happened, and gets a gate rather than a paragraph
  • Built to ship production code, not proof-of-concepts